PMXT Behavioral Clusters v1 — Non-Retail Polymarket Microstructure Dataset

Fill-side behavioral clustering and per-market microstructure metrics for 43,116 Polymarket binary prediction markets over 2026-04-21 to 2026-04-27 UTC. The dataset provides the first public characterisation of non-retail participant behaviour on a decentralised prediction market platform and is released as the empirical foundation for Paper 4 of the Event-Linked Perpetuals research programme.

Corpus

Attribute	Value
Markets	43,116
Unique addresses (aggregate only)	77,203
Total fills	13,356,931
Source	Polymarket CTFExchange OrderFilled events, Polygon mainnet
Block range	86,008,447 – 86,107,178
Date range	2026-04-21 to 2026-04-27 UTC
Pipeline run	2026-05-11

Feature-tier distribution

Tier	Addresses	% of total	Notional	% of total
whale-tier (≥ $1M notional)	68	0.1%	$184M	28.0%
high-frequency-operator	2,952	3.8%	$155M	23.5%
power-trader	6,738	8.7%	$197M	29.9%
active-retail	2,062	2.7%	$70M	10.6%
high-breadth-operator	2,025	2.6%	$7M	1.1%
episodic-retail	63,358	82.1%	$45M	6.8%

Top 3 tiers (12.6% of addresses) control 81.4% of notional volume.

k-means k=5 exploratory partition

Cluster	Archetype	Addresses	Notional share
C1	fill-MM	16,786	46.0%
C2	fill-LP	13,626	45.1%
C0	SPECIALIST	13,775	7.5%
C3	RETAIL	20,033	0.8%
C4	RETAIL	14,733	0.6%

Note: DBSCAN yielded 1 cluster (unimodal data); k-means k=5 is an exploratory fallback (silhouette = 0.227). Feature-tier classification is the recommended primary stratification.

Three-gate verdict

Gate	Result
G-FILL	PASS — 13.36M fills attributed via eth_getLogs
G-QUOTE-LIFE	FAIL universal — off-chain CLOB; no quote lifecycle data
G-BOOK	PASS partial — market-level best_bid/best_ask only

Files

File	Description
data/per_cluster_summary.json	k-means cluster centroids, 95% CI, archetype labels
data/per_cluster_aggregates.parquet	Cluster-level totals and summary statistics (5 rows)
data/per_address_tiers.parquet	Address-level tier and cluster assignments (77,203 rows)
data/tier_kmeans_crosstab.json	6 tiers × 5 clusters cross-tabulation
data/tier_sensitivity.json	Tier populations at P90/P95/P99 threshold variants
data/per_market_microstructure.parquet	43,116 markets × 28 metrics (PR, TS, OI, VPIN, SCI, Kyle's λ)
data/per_market_ils.parquet	ILS at 4 resolution anchor offsets (6,406 / 43,116 scope_pass)
data/per_market_archetype_share.parquet	Per (market, archetype) volume fraction
data/cluster_microstructure_bilateral_real.json	Spearman ρ, BH-FDR, Mann-Whitney, BCa CI (110 tests, 75 significant)
data/manipulation_patterns.json	Wash-volume candidates and book-depth swings
data/gate_report.json	Three-gate empirical verdict
docs/SCHEMA.md	Column-by-column documentation
docs/METHODOLOGY.md	Full methodology
docs/KNOWN_LIMITATIONS.md	Known limitations (10 items)
DATASHEET.md	Datasheets for Datasets format (Gebru et al. 2018)

Related datasets (PMXT family)

Bundle	DOI	Contents
Bundle 1: pmxt-stylized-facts-v1	10.5281/zenodo.20107449	13,314 markets, SF1–SF9, PMXT v2 archive
Bundle 2: pmxt-counterfactual-replay-v1	10.5281/zenodo.20108387	E2/E3 resolution-zone counterfactuals
Bundle 3: pmxt-behavioral-clusters-v1	10.5281/zenodo.XXXXXXXX	This dataset

@dataset{nechepurenko2026pmxt_clusters,
  author    = {Nechepurenko, Maksym},
  title     = {{PMXT Behavioral Clusters v1 --- Non-Retail Polymarket Microstructure Dataset}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.XXXXXXXX},
  license   = {CC-BY-4.0}
}